An MIT research team has developed a machine learning facial recognition system that has spontaneously mimicked a key feature of our own brain’s visual processing system, reports MIT News.
The researchers’ aim was to develop a facial recognition system that could learn to construct what are called “invariant” representations of faces – images of faces that our brains create for recognition that aren’t based on specific angles or rotations, but are sort of the essence of a recognized person’s face. In a primate’s brain, current neuroscience suggests that recognition starts in parts of the brain that are specialized to recognize a face from a certain angle, but eventually proceed to this invariant form of recognition.
One of the steps along the way is recognition based on rotation of the face in either direction, and this is the stage that spontaneously appeared in the MIT researchers’ computational model. In other words, as the facial recognition system sought to proceed from the initiation form of recognition based on seeing a face from a certain angle to the ultimate goal of recognizing a face on a fundamental, “invariant” level, it invented an important middle step.
The researchers are quick to caution that they are not recreating human brain processes. The senior author of the researchers’ report, Tomaso Poggio, tells MIT News, “This is not a proof that we understand what’s going on” in the human brain, but it does offer “strong evidence that we are on the right track.” It points to what could be another important model in the ever-advancing field of facial recognition.
Source: MIT News
December 5, 2016 – by Alex Perala